Intel® Distribution of OpenVINO™ Toolkit
Community assistance about the Intel® Distribution of OpenVINO™ toolkit, OpenCV, and all aspects of computer vision-related on Intel® platforms.
6376 Discussions

Workbench failure during INT8 calibration of onnx

Piyush_Khandelwal
1,072 Views

Hi Community,


I have OpenVINO 2021.4.752 installed on ubuntu 20.04 and been using workbench to perform IR and INT8 conversion for onnx models.


I'm trying to perform INT8 calibration for a onnx file but failing in do so during the process.


I'm guessing it has to be something related to the engine input image size which is 1440x384 because I'm able to perform INT8 for 1120x384


Can you please suggest any workaround or is there any limitation to the engine input size.


Attaching Error Log:

[setupvars.sh] OpenVINO environment initialized
[RUN COMMAND] + pot --direct-dump --progress-bar --output-dir /home/workbench/.workbench/models/14/52/job_artifacts --config /home/workbench/.workbench/models/14/52/scripts/int8_calibration.config.json

0%| |00:00

0%| |00:00
5%|▌ |00:00/home/workbench/.workbench/models/14/52/scripts/job.sh: line 70: 7213 Killed pot --direct-dump --progress-bar --output-dir ${ARTIFACTS_PATH} --config ${JOB_BUNDLE_PATH}/scripts/int8_calibration.config.json


thanks and regards,

Piyush

0 Kudos
4 Replies
IntelSupport
Community Manager
980 Views

Hi Piyush,

 

Thanks for reaching out to us.

 

OpenVINO does support networks that work on different input resolutions. You will need to convert the model to IR format then use the .reshape() method to change the input resolution.

 

Anyhow, which specific OpenVINO ONNX model you are using? Just to clarify, are you unable to run INT8 optimization with 1440x384 image size resolution? Then, the killed POT error that you got, is that the models and configs file exported for the INT8 1120x384 resolution?

 

Meanwhile, what are datasets types you are using? If you are using your own dataset, please share the source here for us to verify it on our end.

 

Regards,

Aznie

 


0 Kudos
Piyush_Khandelwal
923 Views

Hi,

 

Thanks for the reply.

 

We have 2 custom ONNX models which expect 1440x384 and 1120x384 image size resolution as input respectively.

I'm unable to convert the higher resolution model. The INT8 optimization step fails to do the conversion.

It does the job for the lower resolution one so I'm guessing the input size is a factor here to do the conversion.

 

Dataset we are using is "Not Annotated Dataset" which is a zip of our 1000 images which I'm afraid I won't be able to share.

 

 

0 Kudos
IntelSupport
Community Manager
769 Views

Hi Piyush,

 

Would you mind sharing your ONNX model for us to further investigate and validate the issue from our end?

 

If possible, you can share the model privately to my email:

noor.aznie.syaarriehaahx.binti.baharuddin@intel.com

 

 

Regards,

Aznie


0 Kudos
IntelSupport
Community Manager
641 Views

Hi Piyush,


Thank you for your question. If you need any additional information from Intel, please submit a new question as this thread is no longer being monitored. 



Regards,

Aznie


0 Kudos
Reply